https://bugs.kde.org/show_bug.cgi?id=525955
Bug ID: 525955
Summary: kwin_wayland segfaults in Compositor::setupLayers
during fullscreen Vulkan gaming, crashes whole session
Classification: Plasma
Product: kwin
Version First 6.7.5
Reported In:
Platform: Other
OS: Linux
Status: REPORTED
Severity: normal
Priority: NOR
Component: generic-crash
Assignee: [email protected]
Reporter: [email protected]
Target Milestone: ---
Summary: kwin_wayland segfaults in Compositor::setupLayers during fullscreen
Vulkan gaming, crashes whole session
Component: generic-crash (maybe platform-drm, not sure which is right — feel
free to move it)
Version: kwin 6.7.5, CachyOS
Was playing Deadlock (fullscreen, Vulkan, running through Proton
GE-Proton10-29) and kwin_wayland just died mid-game. Not a slowdown or a
stutter first, just instantly gone. Screen goes black basically, and literally
everything else on my desktop lost its Wayland connection at the same time —
Plasma, Konsole, KRunner, KCalc, kded, the xdg-desktop-portal services,
kactivitymanagerd, ksecretd, kdeconnectd, all of it. journalctl is just
wall-to-wall "The Wayland connection broke. Did the Wayland compositor die?"
from every single process, all within about 20ms of each other. Both
xdg-desktop-portal-kde and xdg-desktop-portal-gtk hard crashed too.
The game itself (deadlock.exe) also crashed, but about 48 seconds after kwin
did, and its "crash" is just Wine's own SIGSEGV handler catching a dead surface
— so I'm pretty confident that's just fallout from kwin dying underneath it,
not a separate bug in the game.
Setup:
kwin 6.7.5
mesa 26.2.2
AMD GPU (RADV/radeonsi, confirmed from the thread names in the coredump)
Plasma Wayland session
Kernel log for the actual segfault:
kwin_wayland[906]: segfault at 48600000010 ip 00007f068dabc43e sp
00007fffce2efdc0 error 6 in libc.so.6[bc43e,7f068da24000+1b0000] likely on CPU
6 (core 0, socket 0)
That fault address (0x48600000010) isn't a real pointer — nothing on this
system lives anywhere near that range — so this smells like heap corruption
rather than a plain null deref or an OOM kill. error 6 means it was a write to
an unmapped page from userspace.
Backtrace from coredumpctl:
#0 0x00007f068dabc43e n/a (libc.so.6 + 0xbc43e)
#1 0x00007f068dabd88a n/a (libc.so.6 + 0xbd88a)
#2 0x00007f068e4fbd36 QArrayData::allocate(...) (libQt6Core.so.6 + 0x2fbd36)
#3 0x00007f0690fb875e n/a (libkwin.so.6 + 0x1b875e)
#4 0x00007f0690fb8d4d n/a (libkwin.so.6 + 0x1b8d4d)
#5 0x00007f0690fb035f KWin::Compositor::setupLayers(...) (libkwin.so.6 +
0x1b035f)
#6 0x00007f0690fb27c4 KWin::Compositor::composite(RenderLoop*) (libkwin.so.6 +
0x1b27c4)
#7 0x00007f068e40f13b n/a (libQt6Core.so.6 + 0x20f13b)
#8 0x00007f0691019ec7 KWin::RenderLoop::frameRequested(RenderLoop*)
(libkwin.so.6 + 0x219ec7)
#9 0x00007f068e406039 QObject::event(QEvent*) (libQt6Core.so.6 + 0x206039)
#10 0x00007f068f50642b QApplicationPrivate::notify_helper(...)
(libQt6Widgets.so.6 + 0x10642b)
#11 0x00007f068e395ee8 QCoreApplication::notifyInternal2(...) (libQt6Core.so.6
+ 0x195ee8)
#12 0x00007f068e5a1a2a QTimerInfoList::activateTimers() (libQt6Core.so.6 +
0x3a1a2a)
#13 0x00007f068e5ae5a5 QEventDispatcherUNIX::processEvents(...)
(libQt6Core.so.6 + 0x3ae5a5)
#14 0x00007f068f1f8a22 QUnixEventDispatcherQPA::processEvents(...)
(libQt6Gui.so.6 + 0x7f8a22)
#15 0x00007f068e3a4c54 QEventLoop::exec(...) (libQt6Core.so.6 + 0x1a4c54)
#16 0x00007f068e3999e4 QCoreApplication::exec() (libQt6Core.so.6 + 0x1999e4)
#17 0x000055a3cd64e007 n/a (kwin_wayland + 0x3f007)
#18 0x00007f068da27d0e n/a (libc.so.6 + 0x27d0e)
#19 0x00007f068da27e4b __libc_start_main (libc.so.6 + 0x27e4b)
#20 0x000055a3cd653ea5 n/a (kwin_wayland + 0x44ea5)
Everything else in the coredump is just idle RADV/gallium worker threads
sitting in pthread_cond_wait, so nothing weird happening on other threads at
the time — the crash is squarely in the main thread's compositing path, right
when it's setting up output layers for a frame.
Steps to reproduce: honestly not 100% sure how to trigger it on demand — it
happened during normal play, no alt-tabbing or resolution change that I
noticed. Fullscreen Vulkan + Proton seems to be the common thread. Will update
if I figure out something more reliable.
Happy to attach the actual core dump or run more gdb/coredumpctl stuff against
it if it'd help — I've still got it on disk.
--
You are receiving this mail because:
You are watching all bug changes.
